cadamcat: 1x A100-80GB vs 2x RX 7900 XT TP=2 batch-1 decode speedup ratio (vLLM)
1.4 x (A100 tok/s vs dual-7900XT tok/s, batch-1 decode, 500-32K ctx ladder)
Fields
id
benchmarks/cadamcat-a100-vs-dual-7900xt-decode-speedupMetric
speedup_ratio
Value
1.4
Unit
x (A100 tok/s vs dual-7900XT tok/s, batch-1 decode, 500-32K ctx ladder)
Model
gemma-4-12B-it (dense) + gemma-4-26B-A4B (MoE)
Methodology
THIRD-PARTY MEASURED (cadamcat dual-radeon-vllm, a100-vs-two-radeons). vLLM batch-1 decode, 11 context depths 500->32K, 2 rounds/cell. Stock ladder (no patches both platforms) yields A100 leading 1.36x-1.44x; the nominal BW ratio is 1.27x (2039 vs 2x800 GB/s). Fit r2>=0.9999.
Measured by
third_party
Reproducible
yes
Thermal state
steady
notes
Headline ratio on STOCK ladder = 1.36-1.44x A100 over dual 7900 XT (reported as ~1.40x mean). The second 7900 XT card adds only 1.11-1.23x decode (bandwidth-bound step + all-reduce). Prefill is the opposite: second card buys 1.56-1.87x. Compares two consumer cards vs one used-datacenter A100 - the 'buy another one' decision.
Notes
Headline ratio on STOCK ladder = 1.36-1.44x A100 over dual 7900 XT (reported as ~1.40x mean). The second 7900 XT card adds only 1.11-1.23x decode (bandwidth-bound step + all-reduce). Prefill is the opposite: second card buys 1.56-1.87x. Compares two consumer cards vs one used-datacenter A100 - the 'buy another one' decision.